Aran Khanna

AI Engineer & Researcher

My Projects include Apps I've built, Code I've written and Research I've done

If you're interested in working on a project with me you can check out my Resume and Contact Me →


Apps

Code

Research

Full Products and Apps I've released

Reserved.ai

Reserved.ai is a platform that automates purchasing, renewing and exchanging your AWS Reserved Instances: Optimizing cloud savings so you can focus on Building

The Money Trail

A Chrome Browser Extension that aggregates and allows you to explore all transactions shared by users on their Venmo Profile, to discover patterns in their spending habits.

The Marauders Map

A now defunct Chrome Browser Extension that aggregated and displayed location data shared by your friends via Facebook Messenger.

Open source code projects I've worked on

Apache MXNet

Contributed to the Apache MXNet project. Building capabilties in the deep learning framework to allow for reliable and performant deep network deployment on IoT, mobile and edge devices.

Android Location Tracker

Code for a server that automatically collects a users Android location history and exposes and API to query it, along with frontend code for a map to display location information in the browser.

Sonic Bar Code

Code from HackMIT 2013 for a server to create ultra-sonic "bar codes" associated with URLS and an iOS app capable of discovering them.

Pfast and Pfurious

Code to continuously scrape shuttle location data for the Harvard shuttles from the web app Trans-Loc.

Seam Carving

Code from my CS51 final project to resize images in a content aware manner using a seam carving algorithm.

See all my open source code on Github...

Research I've published in machine learning, digital privacy and security

Stochastic Activation Pruning for Robust Adversarial Defense

A paper presented at NIPS 2017 in the Machine Deception Workshop introducing a technique to allow pre-trained networks to become less suseptible to adversarial attacks by using a novel stochatic "dropout" procedure applied to activation tensors at prediction time, which maintains model accuracy while adding robustness to misclassifying adversarially preturbed inputs.

Presented at NIPS 2017 Machine Deception Workshop

Tensor Contraction & Regression Networks

A paper presented at NIPS 2017 and winner of best poster in the MLTrain Workshop, showing that by replacing flatten and fully connected layers in convolutional neural networks with novel tensor contraction and regession "layers", we can expliot the natural multi-linear structure of activation tensors and reduce the number of parameters needed to achieve state of the art accuracy in popular image classification architectures, including Resnet and VGG.

Presented at NIPS 2017 MLTrain Workshop

StrassenNets: Deep Learning With a Multiplication Budget

A paper presented at NIPS 2017 in the MLTrain Workshop, introducing the use of learnable sum product networks in place of general matrix multiply operations in convolutional neural networks, allowing for a significant reduction in the number and precision of multiply accumulate operations needed to run accurate prediction with these models.

Presented at NIPS 2017 MLTrain Workshop

Tensor Contraction Layers for Parsimonious Deep Nets

A paper presented at CVPR 2017 in the Tensor Methods for Computer Vision Workshop, introducing a new deep network "layer" based on applying the Tucker tensor contraction process to activation tensors. We find that this allows for significant dimensionality reduction of the activation tensor, yielding more efficient model representations without impacting accuracy.

Presented at CVPR 2017 Tensor Methods for Computer Vision Workshop

Venmo’ed: Sharing Your Payment Data With the World

A paper edited by Latanya Sweeney published in Harvard's Technology Science Journal highlighting how Venmo's default privacy settings cause users to inadvertantly publicly share a large amount of private data.

Originally published in Technology Science

Facebook's Privacy Incident Response: a study of geolocation sharing on Facebook Messenger

A paper edited by Latanya Sweeney published in the inaugural issue of Harvard's Technology Science Journal highlighting issues with Facebook's response to a privacy incident. Covered by Boston.com, Forbes, The Washington Post and over 400 other global publications.

Originally published in Technology Science